removing calls to kernel_name_shortener in mongo

shortening now in profile Signed-off-by: JoseSantosAMD <josantos@amd.com> [ROCm/rocprofiler-compute commit: 4e51c122d1]
2023-08-08 12:06:39 -05:00
parent 2765d923a0
commit a712bcef4b
2 changed files with 202 additions and 105 deletions
@@ -53,6 +53,8 @@ from common import (

 from common import getVersion

+cache = dict()
+
 ################################################
 # Helper Functions
 ################################################
@@ -260,7 +262,76 @@ def mongo_import(args, profileAndImport):
    csv_converter.convert_folder(connectionInfo, Extractionlvl)
    print("-- Complete! --")

+def kernel_name_shortener(df, cache, level):
+    if level >= 5:
+        return df

+    columnName = ""
+    if "KernelName" in df:
+        columnName = "KernelName"
+    if "Name" in df:
+        columnName = "Name"
+
+    if columnName == "KernelName" or columnName == "Name":
+        # loop through all indices
+        for index in df.index:
+            original_name = df.loc[index, columnName]
+            if original_name in cache:
+                continue
+
+            # cache miss, add the shortened name to the dictionary
+            new_name = ""
+            matches = ""
+
+            names_and_args = re.compile(r"(?P<name>[( )A-Za-z0-9_]+)([ ,*<>()]+)(::)?")
+
+            # works for name Kokkos::namespace::init_lock_array_kernel_threadid(int) [clone .kd]
+            if names_and_args.search(original_name):
+                matches = names_and_args.findall(original_name)
+            else:
+                # Works for first case  '__amd_rocclr_fillBuffer.kd'
+                # remove .kd and then parse through original regex
+                first_case = re.compile(r"([^\s]+)(.kd)")
+                Mod_name_and_args = re.compile(r"(?P<name>[( )A-Za-z0-9_]+)([ ,*<>()]*)")
+                interim_name = first_case.search(original_name).group(1)
+                matches = Mod_name_and_args.findall(interim_name)
+
+            current_level = 0
+            for name in matches:
+                ##can cause errors if a function name or argument is equal to 'clone'
+                if name[0] == "clone":
+                    continue
+                if len(name) == 3:
+                    if name[2] == "::":
+                        continue
+
+                if current_level < level:
+                    new_name += name[0]
+                # closing '>' is to be taken account by the while loop
+                if name[1].count(">") == 0:
+                    if current_level < level:
+                        if not (current_level == level - 1 and name[1].count("<") > 0):
+                            new_name += name[1]
+                    current_level += name[1].count("<")
+
+                curr_index = 0
+                # cases include '>'  '> >, ' have to go in depth here to not lose account of commas and current level
+                while name[1].count(">") > 0 and curr_index < len(name[1]):
+                    if current_level < level:
+                        new_name += name[1][curr_index:]
+                        current_level -= name[1][curr_index:].count(">")
+                        curr_index = len(name[1])
+                    elif name[1][curr_index] == (">"):
+                        current_level -= 1
+                    curr_index += 1
+
+            cache[original_name] = new_name
+            if new_name == None or new_name == "":
+                cache[original_name] = original_name
+
+        df[columnName] = df[columnName].map(cache)
+
+    return df
 ################################################
 # Roofline Helpers
 ################################################
@@ -768,6 +839,7 @@ def main():
    # PROFILE MODE
    ##############
    if args.mode == "profile":
+        Extractionlvl = 3 #args.extraction_level
        print("Resolving rocprof")
        resolve_rocprof()
        # Cannot access parent directories
@@ -805,11 +877,43 @@ def main():
            roof_setup(args, my_parser, VER)
            # Generate roofline
            roofline_only(args.path, args.device, args.sort, args.mem_level, args.kernel_names, args.verbose)
+            #demangle
+            for file in os.listdir(args.path):
+                if file.endswith(".csv"):
+                    try:
+                        fileName = file[0 : file.find(".")]
+                        # Only shorten KernelNames if instructed to
+                        if Extractionlvl < 5:
+                            t1 = pd.read_csv(
+                                os.listdir(args.path) + "/" + file,
+                                on_bad_lines="skip",
+                                engine="python",
+                            )
+
+                            t2 = kernel_name_shortener(t1, cache, level=Extractionlvl)
+                    except pd.errors.EmptyDataError:
+                        print("Skipping empty csv " + file)

        # Profile only
        else:
            print("\n-------------\nProfile only\n-------------\n")
            omniperf_profile(args, VER)
+            #demangle
+            for file in os.listdir(args.path):
+                if file.endswith(".csv"):
+                    try:
+                        fileName = file[0 : file.find(".")]
+                        # Only shorten KernelNames if instructed to
+                        if Extractionlvl < 5:
+                            t1 = pd.read_csv(
+                                os.listdir(args.path) + "/" + file,
+                                on_bad_lines="skip",
+                                engine="python",
+                            )
+
+                            t2 = kernel_name_shortener(t1, cache, level=Extractionlvl)
+                    except pd.errors.EmptyDataError:
+                        print("Skipping empty csv " + file)

    ##############
    # DATABASE MODE